AI has existed for decades, but its usage in the marketing scene has remained relatively subtle and unnoticeable to most people. After 2020, advanced AI has become mainstream, and almost everyone feels its impact somehow. The first big wave of generative AI came with the launch of ChatGpt 3.5 in 2022, followed soon after by Google’s Bard, which has since upgraded to Gemini.
Recently, Google released an AI tool based on its Gemini technology, optimized for use in environments with limited resources, such as a laptop or a small cloud infrastructure. According to Google, the new Gemma tool can create Chabots, generate content, and do everything other language models, such as Gemini, can do.
It Is an Open-License Model
This tool has two versions: a two billion and a seven billion parameter version. The parameter value is a measure of the tool’s capabilities. While the two billion parameter version can still handle complex tasks, the seven billion parameter version handles even more complex tasks, understands language much better, and can generate more sophisticated responses. However, it requires much more resources for training and running, so an organization’s choice will largely depend on the available resources.
Anyone can use the tool for commercial suh as campaigns and non-commercial purposes under its open license. An open license is a legal agreement where a creator allows other people to use their creation, shape it, and modify it but places limitations on how it can be used. With AI technology posing serious safety threats, Google has placed restrictions against using this tool for malicious purposes.
“The reason behind the creation of this tool was to have a lightweight enough tool for usage in almost any environment, allowing it to get into the hands of as many end users as possible. It’s the kind of tool SEOs have always dreamed of having in their hands,” says Washington, DC, marketing expert Seth Price.
Its Parameters
The tool has a vocabulary of 250k tokens, with comparable models having a meager 32k tokens. This upper hand allows it to process inputs and generate outputs much faster than other models available in the industry. Its embedding weights are also one of a kind at 750 million.
Embedding weights are the parameters used to map words with meanings and relationships with the outputs. For end users, higher embedding weights mean more accuracy, relevance, and contextually accurate responses. However, it is important to note that the effectiveness of a tool largely depends on the training it receives. The good thing is its machine learning capabilities help it get better with use.
Safety Concerns
The concern for most users of new technologies is their safety. According to Google, Gemma is designed to be safe from the bottom up and users should not be concerned when deploying it. Its training data has been subjected to extensive filtering to ensure it removes personal or sensitive data.
Reinforced learning from human feedback in its creation ensures that the tool doesn’t go outside responsible behavior parameters. It has also been subjected to extensive testing, further manual debugging, and automated testing to determine and eliminate its capability for dangerous and unwanted activity.
The good thing is its open license allows a user to fit their industry. So, you can work on enhanced safety through customization if you are concerned about its use. The best place to get information on customization and safety improvement is Google’s responsible generative AI tool kit. The tool kit offers guidance on safety classification and debugging, which helps you investigate its behavior, and general guidance on the best practices for building based on Google’s experience and applications.
